A new fast match for very large vocabulary continuous speech recognition

A novel fast match for very large vocabulary continuous speech recognition in which phonetic transcriptions are scored at a cost of one floating-point operation per phone is presented. The algorithm uses a table of estimates of phone scores and durations derived by searching a relatively small phonetic graph that imposes triphone phonotactic constraints on phone strings. The computational cost of the preprocessor (which is essentially independent of the size of the dictionary) is a few times real time on a moderately powerful work station.<>